Automatic optimization with uncertainty aware neural networks

Abstract

A method for automatic optimization of a system includes randomly generating a plurality of input parameter configurations for the system. Using a trained neural network, a plurality of throughputs of the system are simulated using each of the randomly generated input parameter configurations. Each of the randomly generated input parameter configurations are scored based on the simulated throughputs and data stored in a training database. An input parameter configuration is selected from the randomly generated plurality of input parameter configurations based on the scoring. The selected input parameter configuration is sent to an actuator for executing the system using the selected input parameter configuration.

What is claimed is: 
     
         1 . A method for automatic optimization of a system, the method comprising:
 randomly generating a plurality of input parameter configurations for the system;   simulating, using a trained neural network, a plurality of throughputs of the system using each of the randomly generated input parameter configurations;   scoring each of the randomly generated input parameter configurations based on the simulated throughputs and data stored in a training database;   selecting an input parameter configuration from the randomly generated plurality of input parameter configurations based on the scoring; and   sending the selected input parameter configuration to an actuator for executing the system using the selected input parameter configuration.   
     
     
         2 . The method of  claim 1 , further comprising measuring a throughput of the system executed using the selected input parameter configuration. 
     
     
         3 . The method of  claim 2 , wherein the measured throughput of the system and the selected input parameter configuration are stored in the training database. 
     
     
         4 . The method of  claim 3 , wherein the data stored in the training database is used to train the neural network. 
     
     
         5 . The method of  claim 4 , wherein the neural network further comprises:
 a prediction branch that simulates the throughput of the system using a parameter configuration from the randomly generated input parameter configurations; and   an uncertainty branch that determines a confidence level of the simulated throughput of the system using the parameter configuration from the randomly generated input parameter configurations.   
     
     
         6 . The method of  claim 5 , wherein the uncertainty branch of the neural network estimates an error associated with each of the simulated throughputs. 
     
     
         7 . The method of  claim 6 , wherein the error associated with each of the simulated throughputs is estimated by comparing the simulated throughputs to data stored in the training database. 
     
     
         8 . The method of  claim 7 , wherein training the neural network comprises:
 reducing the estimated error associated with each of the simulated throughputs; and   reducing losses associated with the prediction branch of the neural network.   
     
     
         9 . The method of  claim 7 , wherein the scoring of the randomly generated input parameter configurations is based on the simulated throughputs and the estimated error associated with each of the simulated throughputs. 
     
     
         10 . The method of  claim 1 , wherein the system is a physical system interacted in a wet lab or a chemical facility, and wherein the input parameter configurations include a configuration of a plurality of components that are part of a chemical reaction. 
     
     
         11 . The method of  claim 1 , wherein the system is an operating system, and wherein the input parameter configurations include a configuration of one or more of the following parameters of the operating system: worker_connections, payload_size, keepalive_timeout, open_file_cache, num_parallel_connections, tcp_nopush. 
     
     
         12 . The method of  claim 11 , wherein the throughput of the operating system is based on a number of requests per second handled by the operating system and power consumption of the operating system. 
     
     
         13 . The method of  claim 1 , wherein scoring each of the randomly generated input parameter configurations further comprises computing a weighted combination of probability improvement and a minimal distance between a sampled point and training data. 
     
     
         14 . A computer system programmed for automatic optimization of a system, the computer system comprising one or more hardware processors configured by code stored in memory to provide for execution of the following steps:
 randomly generating a plurality of input parameter configurations for the system;   simulating, using a trained neural network, a plurality of throughputs of the system using each of the randomly generated input parameter configurations;   scoring each of the randomly generated input parameter configurations based on the simulated throughputs and data stored in a training database;   selecting an input parameter configuration from the randomly generated plurality of input parameter configurations based on the scoring; and   sending the selected input parameter configuration to an actuator for executing the system using the selected input parameter configuration.   
     
     
         15 . A tangible, non-transitory computer-readable medium having instructions thereon, which, upon being executed by one or more processors, provides for execution of the following steps:
 randomly generating a plurality of input parameter configurations for the system;   simulating, using a trained neural network, a plurality of throughputs of the system using each of the randomly generated input parameter configurations;   scoring each of the randomly generated input parameter configurations based on the simulated throughputs and data stored in a training database;   selecting an input parameter configuration from the randomly generated plurality of input parameter configurations based on the scoring; and   sending the selected input parameter configuration to an actuator for executing the system using the selected input parameter configuration.